🚀 go-pugleaf

              DEPARTMENT OF COMPUTER SCIENCE
                COLORADO STATE UNIVERSITY

          BARNEY'S MONDAY AFTERNOON CLUB SEMINAR

      Speaker:	Dr. Kiriakos N. Kutulakos
		Dept of Computer Sciences
	        University of Rochester, Rochester, NY

	        Dr. Kutulakos is a candidate for a tenure 
	        track position with us.

         Date:	Friday, May 3, 1996 [Note:  Not a Monday]

	 Time:	3:00-4:00 p.m.

         Place:	100 USC

         INTERACTIVE AND AUTONOMOUS 3D MODELING
                 FROM LIVE VIDEO IMAGES

An increasing number of applications, from virtual reality to 
reverse engineering, require building and interacting with 
realistic models of complex 3D objects.  The availability of 
low-cost, high-performance 3D rendering and video acquisition 
hardware promises a new generation of multimedia tools and 
techniques that use live video as input, interact with the 
user or the user's environment, and are robust and cost-
effective.

I will discuss two complementary approaches to the video-
based modeling problem.  Both are specifically designed to 
impose as few restrictions as possible on the shape of the 
object being modeled.  The first, called "3D Stenciling," 
exploits the modeling system's ability to interact with its 
user via a hand-held pointer and an augmented reality inter-
fact.  The second avoids user interaction altogether by 
attempting to build global surface models automatically.

I will focus on two central aspects of 3D Stenciling:  
(1) providing direct visual feedback to the user by
registering the interactively reconstructed graphical 
model with live video of the object being modeled, and
(2) eliminating the need for camera calibration by form-
ulating the video overlay and 3D modeling problems within 
an affine framework.

I will describe an approach for automatically constructing 
global models for objects with arbitrary curved surfaces.  
In this case, automatic control of the camera's viewpoint 
becomes crucial even when constructing models of simple 
non-convex objects (e.g., pipes).  I will outline a
provably correct strategy for controlling the camera's 
viewpoint in real time that maintains a well-defined 
(and very simple) geometric relationship between the 
viewpoint's motion and the object's deforming silhouette.

The two approaches present opposite ends of a spectrum 
of video-based modeling techniques that balance user 
interaction and autonomy to build models of physical 3D 
objects.